kaijieshi7 / dynamic-convolution-pytorch Goto Github PK

I am really interested in a fantastic Dynamic Convolution and I intend to use it in my research but I can not figure out how to use it.
I can not determine the relation between in_planes , out_planes, groups, and other parameters and how it affects the performance.
hope you will provide more clarification about the code and parameters' relationship to each other and to the input.

self.bias = nn.Parameter(torch.Tensor(K, out_planes))
should be changed to
self.bias = nn.Parameter(torch.zeros(K, out_planes))

当替换分组卷积时，weight的维度和group的维度不对应。
RuntimeError: Given groups=1024, expected weight to be at least 1024 at dimension 0, but got weight of size [32, 32, 3, 3] instead
aggregate_weight = torch.mm(softmax_attention, weight).view(-1, self.in_planes, self.kernel_size, self.kernel_size)
这一行应改成下面这样
aggregate_weight = torch.mm(softmax_attention, weight).view(-1, self.in_planes//self.groups, self.kernel_size, self.kernel_size)

加入动态卷积之后，shape就变了，原来的ResNet预训练模型就不能用了，请问大佬有动态卷积的ResNet50预训练模型吗？

大佬你好！
for idx_batch, (inputs, targets) in enumerate(trainloader):
这句for循环不知道为什么不执行

My image size is 112,112 and I got the below error. Any solution?

Line number
ConvLayers/dynamic_conv.py", line 176, in forward

_x = x.view(1, -1, height, width)# 变化成一个维度进行组卷积
RuntimeError: view size is not compatible with input tensor's size and stride (at least one dimension spans across two contiguous subspaces). Use .reshape(...)_

Dynamic conv3d中好像没有对卷积核参数做初始化，请问这是为什么呢？这样会有什么影响吗？

你好，请问我将动态卷积加入resnet，预训练模型不匹配，应该如何加载预训练权重的问题呀

当stride 为2的时候，这个是不是不能用了

请问您有自己实验了吗？速度和精度能达到论文那个水平吗

感谢你的复现并开源，请问你有试过你写的动态卷积的速度吗，速度如何？谢谢！

Thank you for providing this reproduction!

I have a question on the grouped convolution: in this line you use the grouped convolution to solve the mini-batch training problem.

Could we use the torch.Tensor.expand to replace the grouped convolution, like:

weight_prime = weight.expand(K, weight.shape[0], weight.shape[1], weight.shape[2], weight.shape[3])
weight = torch.mm(softmax_attention, weight_prime).view(-1, x.shape[1], self.kernel_size, self.kernel_size)

In this way, we might aggregate the attention weight and the convolution weight together. However, this may cause another problem. If batch size ($\mathcal{B}$) is larger than 1, the attention weight would be a matrix with $\mathcal{B} \times K$, I think we can use torch.mean(attention_weight, dim=0) or torch.max(attention_weight, dim=0) since they are calculated within the batch, in which the range is very close.

I am not sure whether this calculation is equivalent to the line :)

我想请教下各位大佬，我在调试时发现每个卷积核的权重都是0.25，想问这是什么原因呢？怎么解决呢？

feels like K = out channels? is it right?

作者您好，感谢您对论文的复现！

程序中关于温度问题我还不太明白，期待您的回复：
1.不同的温度值对于模型的训练有什么影响呢？在程序的哪个部分可以体现呢？
2.如果想要使训练温度保持恒定，即不使用temperature annealing，在程序中需要做哪些改动呢？

再次表示感谢！祝您一切顺利！！

I don't observe any accuracy boost for dy_resnet18 compared to raw_resnet18 model on CIFAR10 and CIFAR100 datasets.
I used default hyper-parameters for the training and observed.
CIFAR10: dy 86.01%, raw 87.75%
CIFAR100: dy 54.59%, raw 56.52%.
All the experiments were bellow my expectations (>90% for CIFAR10, >70% for CIFAR100).
How to take advantage of dynamic convs?

很感谢你的分享，但是我在调试过程中发现当输入为（128,1,20,30,30）这个tensor的通道为1时，代码会报错，
File "D:\anaconda3\envs\py38\lib\site-packages\torch\nn\modules\conv.py", line 557, in init
super(Conv3d, self).init(
File "D:\anaconda3\envs\py38\lib\site-packages\torch\nn\modules\conv.py", line 83, in init
self.reset_parameters()
File "D:\anaconda3\envs\py38\lib\site-packages\torch\nn\modules\conv.py", line 86, in reset_parameters
init.kaiming_uniform_(self.weight, a=math.sqrt(5))
File "D:\anaconda3\envs\py38\lib\site-packages\torch\nn\init.py", line 381, in kaiming_uniform_
fan = _calculate_correct_fan(tensor, mode)
File "D:\anaconda3\envs\py38\lib\site-packages\torch\nn\init.py", line 350, in _calculate_correct_fan
fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
File "D:\anaconda3\envs\py38\lib\site-packages\torch\nn\init.py", line 282, in _calculate_fan_in_and_fan_out
receptive_field_size = tensor[0][0].numel()
IndexError: index 0 is out of bounds for dimension 0 with size 0

如题，尝试了很多种方式都不支持动态权重,您有什么好的建议吗？

在打印的网络中并没有看到动态生成的weight和bias，这样是可以训练的嘛？（ps:此前在训练层注意力中，也用到了beta=nn.parameter,但是定义的参数beta值一直都没有改动！！！）。能问下您的pytorch版本嘛？

您好，感谢您的代码！我在运行前将cifar10放入新建的data文件夹中，然后运行main时解压成了如图文件。之后就产生了没有model定义的报错，如图所示。请问应该如何正确运行呢？如果能解答，万分感谢！！

dynamic convolution能否使用在时间序列问题上，我想用dynamic convolution来提取时间特征，感谢分享

一直限制attention的输出大小，这是warmup的手法吗？为什么是加在attention后面而不是卷积后面呢？